Updates

vendor/agent-harness/lib/catalog.js

@@ -14,7 +14,7 @@ const CATALOG = [
     mmproj: 'MMPROJ_QWEN3_5_0_8B_MULTIMODAL_Q8_0',
     minRamGb: 4,
     approxDownloadGb: 0.7,
-    ctxSize: 8192,
+    ctxSize: 16384,
   },
   {
     id: 'qwen3.5-2b',
@@ -26,7 +26,7 @@ const CATALOG = [
     mmproj: 'MMPROJ_QWEN3_5_2B_MULTIMODAL_Q8_0',
     minRamGb: 6,
     approxDownloadGb: 1.6,
-    ctxSize: 8192,
+    ctxSize: 16384,
   },
   {
     id: 'qwen3.5-4b',
@@ -155,7 +155,7 @@ const CATALOG = [
     vision: false,
     minRamGb: 8,
     approxDownloadGb: 1.2,
-    ctxSize: 8192,
+    ctxSize: 16384,
   },
   {
     id: 'qwen3-4b',
@@ -315,6 +315,9 @@ const COMPACT_TOOL_ALLOW = [
   'memory_search',
   'memory_get',
   'memory_write',
+  'todo_write',
+  'task',
+  'update_goal',
   'capability_status',
   'qvac_runtime_state',
   'qvac_system_resources',
@@ -332,6 +335,54 @@ function filterToolsForModel(defs, idOrConstant) {
   return filtered.length ? filtered : list;
 }
 
+function catalogId(idOrConstant) {
+  const e = findCatalogEntry(idOrConstant);
+  return String((e && e.id) || idOrConstant || '').toLowerCase();
+}
+
+function isQwenReasoningModel(idOrConstant) {
+  const id = catalogId(idOrConstant);
+  return /qwen3(?:\.\d+)?/.test(id) || /qwen3[._-]/.test(String(idOrConstant || '').toLowerCase());
+}
+
+// Qwen3.5's chat template leaves thinking off unless the reasoning channel is
+// enabled. Cap it so HUD thinking still streams, then the addon force-closes
+// <think> and the model can emit the tool XML / spoken reply.
+function reasoningBudgetForModel(idOrConstant) {
+  if (!isQwenReasoningModel(idOrConstant)) return null;
+  const id = catalogId(idOrConstant);
+  if (/0\.8b|0\.6b|600m/.test(id)) return 128;
+  if (/1\.7b|(?:^|-)2b/.test(id)) return 192;
+  if (/(?:^|-)4b/.test(id)) return 320;
+  if (/(?:^|-)[89]b/.test(id)) return 448;
+  return 320;
+}
+
+// Slightly below Qwen thinking's 0.6 so compact models still follow the
+// tool XML, but high enough to avoid the greedy repeat loops of temp 0.3.
+function compactGenerationParams(idOrConstant) {
+  if (!isCompactToolModel(idOrConstant)) return {};
+  return { temp: 0.55 };
+}
+
+function generationParamsForModel(idOrConstant, extra) {
+  const base = Object.assign(
+    {
+      repeat_penalty: 1.15,
+      frequency_penalty: 0.15,
+    },
+    compactGenerationParams(idOrConstant)
+  );
+  const budget = reasoningBudgetForModel(idOrConstant);
+  if (budget != null) {
+    base.reasoning_budget = budget;
+    base.predict = Math.min(768, Math.max(384, budget + 288));
+  } else {
+    base.predict = isCompactToolModel(idOrConstant) ? 384 : 640;
+  }
+  return Object.assign({}, base, extra || {});
+}
+
 module.exports = {
   CATALOG,
   FALLBACK_LLM_IDS,
@@ -346,5 +397,9 @@ module.exports = {
   isCompactToolModel,
   compactToolAllowlist,
   filterToolsForModel,
+  isQwenReasoningModel,
+  reasoningBudgetForModel,
+  compactGenerationParams,
+  generationParamsForModel,
   catalogLabel,
 };